Enhancing the quality of photographs is a highly subjective process and depends on users' preferences. Hence, it is often more desired to let users choose their own best from a set of diverse and adjustable enhanced images with astounding quality. However, a system that can satisfy this requirement has not yet been established. While classical algorithms blindly enhance an image by filtering, recent intelligent enhancement systems can only do it with limited styles through learning from a set of single expert-retouched (ER) images. To fill this void, we propose a novel framework, Diverse and adjustable Versatile Image Enhancer (DaVIE), that learns from multiple ER images simultaneously. Thereby, it can output diverse results without being bound to a specific enhancement style while allowing users to freely adjust the level of enhancement. For ease of diversity, we adopt a variational auto-encoder (VAE) that learns stochastic distribution of enhancement styles. By using the VAE, the proposed model provides diversely enhanced images. To establish better control in terms of enhancement level, we propose a more general form of adaptive instance normalization and loss functions, which can afford even extreme image editing. Through rigorous experiments, we demonstrate that the proposed DaVIE framework yields visually pleasing and diverse results. We also show the proposed model quantitatively outperforms existing methods on the MIT-Adobe-5K dataset. Furthermore, through a strict user-study, we show that the users consider the qualities of ER images and machine-retouched images to be similar, with about 35% selection probability for DaVIE enhanced images.
|Number of pages||14|
|Publication status||Published - 2021|
Bibliographical noteFunding Information:
This work was supported in part by the National Research Foundation of Korea (NRF) through the Korea Government (Ministry of Science and ICT, MSIT) under Grant NRF-2020R1A2C3011697, and in part by the Yonsei University Research Fund of 2021 under Grant 2021-22-0001.
© 2013 IEEE.
All Science Journal Classification (ASJC) codes
- Computer Science(all)
- Materials Science(all)
- Electrical and Electronic Engineering